Obtaining linguistic structure in continuous speech  

Nina Svärd, Pia Nehme & Francisco Lacerda

Fonetik 2002, the XVth Swedish Phonetics Conference, Fysikcentrum, Stockholm, May 29 - 31, 2002. Quartely Progress and Status Report (Department of Speech, Music and Hearing and Centre for Speech Technology, KTH, Stockholm) 44: 9-12.


A simple model that structures continuous speech signals by suggesting word candidates is presented in this paper. To create possible representations of target word candidates a program was written in C++. Text materials in Portuguese and "Svensiska", a nonsense language containing similar phonotactic and prosodic features as Swedish were used in this study. Since the model aims to resemble the language learning process of the infant, the material was based on infant directed speech. The results showed that certain words in the Portuguese material were successfully extracted as being word candidates by the model. However, the results indicated that there were no such patterns to be found in the Svensiska material, which may be attributed to the structure of the Svensiska material. The importance of the relative frequency of sound sequences, their position within the phrase and the variety in the speech signal to the infants' ability to discover patterns and words in the speech signal are discussed.

View pdf-version of paper

PERILUS | All issues | 2002 issue | Previous | Next | Phonetics at Stockholm University